What it is
Sora was OpenAI's video generator. You typed a description, such as "a golden retriever surfing at sunset, filmed from a drone", and a few minutes later it gave you a short video clip that had never been filmed by anyone. Later versions could also start from a photo or an existing clip, and the second generation added sound: speech, sound effects and background noise made to match the pictures.
Sora has been discontinued. OpenAI announced the decision on 24 March 2026. The Sora app and website closed on 26 April 2026, and OpenAI's developer list of retired models shows the Sora 2 models and the Videos API (the service that let other companies' software use Sora) switching off on 24 September 2026, with no replacement listed. OpenAI's help centre says that once any final export window has passed, it will permanently delete the data people created with Sora. OpenAI's original launch pages now carry a line saying the product is no longer available. At the time of writing, OpenAI does not offer a replacement video tool: reports at the time of the announcement, citing The Wall Street Journal, said there were no plans to build Sora's video generation into ChatGPT. An OpenAI spokesperson told Rolling Stone that the Sora research team would continue, focused on "world simulation research to advance robotics". This page therefore describes a product that has ended. If you want to make AI video today, see Runway, Veo and Kling.
Sora had two main lives:
- Sora (2024). OpenAI first showed Sora on 15 February 2024 as a research preview. It could make clips up to a minute long, and it was given only to "red teamers", outside experts in areas like misinformation and bias whose job was to try to make it misbehave, plus a small group of artists and filmmakers. On 9 December 2024 a faster version called Sora Turbo opened to paying ChatGPT Plus and Pro subscribers at sora.com, making clips of up to 20 seconds at up to 1080p resolution (full HD).
- Sora 2 and the Sora app (2025). On 30 September 2025 OpenAI released Sora 2 together with a new invitation-only iPhone app, also called Sora, first in the US and Canada. The app looked and worked much like TikTok: an endless feed of short videos, except that every video was AI-generated. Its headline feature let you record a short video of yourself so that you, or friends you approved, could put your face and voice into generated scenes. OpenAI first called these "cameos" and later renamed them "characters".
How did it work? Sora was a diffusion model, the same family of technique used by most image generators: it starts from random visual noise and cleans it up, step by step, into a picture that matches the prompt. OpenAI combined this with a transformer, the architecture behind ChatGPT, and treated a video as a block of small patches across both space and time, so the whole clip was cleaned up together rather than one frame after another. That is why Sora clips hung together better than earlier AI video, where objects often melted from frame to frame. OpenAI also described video models like Sora as a possible path towards "world simulators", AI that learns how the physical world behaves; that larger idea, and its limits, are covered on the World Models page.
Why it matters
It moved AI video from a curiosity to a mainstream worry and a mainstream toy. Before February 2024, most AI video was a few seconds of wobbling shapes. Sora's first samples, such as a woman walking down a neon-lit Tokyo street at night, looked close enough to real footage that people who work in film, advertising and news took notice all at once. Its competitors, including Runway, Google's Veo and Kling, were measured against it for the next two years.
It showed the cost of consent problems at scale. The Sora app made it easy to put real people, living and dead, into realistic video. Within weeks it faced complaints from Hollywood studios about copyrighted characters, from the family of Martin Luther King Jr. about his image, and from actors about their faces and voices. Each one forced OpenAI to change its rules after launch. Anyone thinking about AI and likeness, meaning a person's face, voice and recognisable identity, can learn from how those arguments played out.
It tested whether labels on AI content work. Every Sora video carried a visible moving watermark and C2PA "content credentials", a hidden label inside the file recording that the video was AI-made (explained on the OpenAI Images page). Within days, websites offering to strip the visible watermark were easy to find. Sora is a clear case study in why a label alone doesn't stop a fake from spreading.
It is a reminder that AI products can disappear. Sora went from preview to discontinued in just over two years, and the Sora app lasted less than seven months. People who built habits, businesses or software around it had to move their work elsewhere. That is worth remembering before you rely on any single AI tool, from any company.
Real-world examples
Air Head: the showcase film and what it took to make it
On 25 March 2024, while Sora was still a closed preview, OpenAI published a set of short films made by artists it had given early access to. The most widely shared was Air Head, by the Toronto production company shy kids: a gentle, funny short about a man whose head is a yellow balloon. It was produced by Sidney Leeder, written and directed by Walter Woodman, and edited by Patrick Cederberg. To many viewers it looked like proof that one prompt could produce a finished film.
A few weeks later, in an interview with the visual effects news site fxguide published on 14 April 2024, Cederberg described how the film was actually made. Sora couldn't keep the balloon consistent from one shot to the next: sometimes it wasn't yellow, and sometimes it had a face drawn on it. The team fixed these problems by hand in Adobe After Effects, isolating and recolouring the balloon and painting out unwanted faces. Camera instructions such as "pan" worked, by his estimate, about six times out of ten. He guessed that the team generated around 300 seconds of footage for every second that made the final cut. The imagery really did come from Sora, but the film also depended on a script, a voiceover, careful editing and traditional effects work. It is a useful benchmark for any "made entirely with AI" claim: ask who did the editing, and how much was thrown away.
The Sora app's first month: likeness, copyright and deepfakes
When the Sora app opened on 30 September 2025, it was an instant hit: CNBC reported one million downloads in under five days. It also set off a series of disputes in a few weeks.
Copyright. Just before launch, The Wall Street Journal reported that OpenAI had told Hollywood studios and agencies that their characters could appear in Sora videos unless they opted out. The feed soon filled with well-known cartoon and game characters. On 3 October 2025, OpenAI's chief executive Sam Altman wrote on his blog that OpenAI would give rights holders "more granular control over generation of characters", closer to the opt-in approach used for people's faces, and would try sharing revenue with rights holders who chose to take part.
Historical figures. Some users made disrespectful videos of Martin Luther King Jr. On 16 October 2025, OpenAI and the King estate issued a joint statement saying that, at the estate's request, OpenAI had paused Sora videos depicting Dr. King while it strengthened guardrails for historical figures, and that authorised representatives of other public figures could ask for their likeness not to be used.
Actors. The actor Bryan Cranston found videos using his face and voice made without his permission. On 20 October 2025, OpenAI, Cranston and the actors' union SAG-AFTRA issued a joint statement: OpenAI said it had strengthened its guardrails on copying the voice and likeness of people who had not opted in, and would work with the union and three talent agencies on protections for performers.
Misinformation. The news-reliability company NewsGuard tested Sora 2 with 20 false claims taken from recent news and reported on 17 October 2025 that it produced convincing videos for 16 of them, 80 percent. By then, 404 Media had already reported, on 7 October, that several websites could remove Sora's visible watermark in seconds.
Even the feature's name was contested. The celebrity video-message company Cameo sued over OpenAI's use of "cameo", and on 24 November 2025 a US federal judge temporarily barred OpenAI from using the word; OpenAI renamed the feature "characters".
Disney's billion-dollar deal, and its collapse
On 11 December 2025, Disney and OpenAI announced a three-year licensing agreement. Sora users would be able to make short videos with more than 200 animated, masked and creature characters from Disney, Marvel, Pixar and Star Wars; a selection of fan videos would appear on Disney+; and Disney would make a 1 billion dollar equity investment in OpenAI. The announcement said the deal did not include any actors' likenesses or voices. It looked like the answer to Sora's copyright problem: licensed characters, with the owner's permission and involvement. Hollywood unions and the Writers Guild criticised it within days.
Just over three months later, on 24 March 2026, the official Sora account posted on X: "We're saying goodbye to Sora." OpenAI gave no detailed reason. News reports linked the decision to the high computing cost of video and to OpenAI's shift of attention towards coding tools and business customers; TechCrunch reported that monthly app downloads had fallen from a peak of 3.3 million in November 2025 to 1.1 million in February 2026. The Disney investment never closed. A Disney spokesperson said the company respected "OpenAI's decision to exit the video generation business" and would keep working with AI platforms. For anyone judging AI partnerships, it is a striking lesson: even a headline deal with the biggest name in entertainment could not keep an expensive product running.
An everyday scenario: an author's lost trailer clips
Picture Priya, a self-published fantasy author. In early 2026 she made a handful of moody Sora clips, a candlelit library and a storm over a harbour, for a 30-second book trailer, and left them sitting in her Sora account. When she hears the app is closing, she follows OpenAI's export steps and downloads her clips, checks that she still has the finished trailer file, and saves her original prompts in a document. Later, when she wants a new trailer for book two, she rewrites those prompts for another video tool. Her lesson: keep your own copies of anything you make with an online AI tool, and keep your prompts, because the tool may not be there next year.
Try it yourself
You can no longer make videos with Sora. These steps help you understand what it could and couldn't do, check AI video you come across, and try a tool that still works. Product names, prices and free allowances change, so check each site's current terms.
- Watch OpenAI's original samples. Open OpenAI's February 2024 page, "Sora: Creating video from text", at openai.com/index/sora/. It is still online. Watch three or four clips all the way through.
- Hunt for the mistakes OpenAI admitted. Further down the same page, OpenAI lists weaknesses such as objects "morphing", people or animals appearing from nowhere, and a chair that doesn't behave like a solid object. Go back to the clips and look for hands, reflections, background people and anything that is picked up or bitten. Spotting these flaws is a skill that helps with every AI video, not just Sora.
- See how a showcase film was really made. Find Air Head on OpenAI's "Sora: first impressions" page (openai.com/index/sora-first-impressions/), then read fxguide's interview "Actually using SORA". Note every step that involved a human: writing, choosing shots, editing, recolouring, removing artefacts.
- If you had a Sora account, check your data now. Read OpenAI's help page "What to know about the Sora discontinuation" for the latest on exports. OpenAI says data will be permanently deleted after any final export window, so don't wait.
- Check a video for content credentials. Go to contentcredentials.org/verify and drop in an image or video file you have downloaded. If it carries C2PA content credentials, the site shows who or what made it. If nothing shows, that proves nothing: labels are often stripped when files are re-saved or uploaded to social media.
- Try a video tool that still works. At the time of writing, Runway offers a free plan with a one-time allowance of credits, and Google offers video generation in its Gemini app (Google says a model called Gemini Omni is replacing Veo there). See the Runway and Veo pages first. Then try a prompt like this one, written in the style Sora users learned worked best, with the subject, setting, camera and mood spelled out: > A slow, steady drone shot drifting over a small stone harbour at dawn. Fishing boats rock gently, gulls circle, mist lies on the water. Soft golden light, muted colours, calm and quiet mood. No people, no text on screen.
- Generate it twice and compare. Like Sora, these tools give a different result each time. Note what stays the same and what changes, and which of your words the tool seems to ignore.
- Label it if you share it. If you post AI video anywhere, say that it is AI-generated, and check the platform's own rules on labelling. Never make realistic video of a real person without their clear permission.
Common mistakes and misunderstandings
Thinking Sora is still available. Many guides, videos and books about Sora are still online and still for sale, and some sites use the Sora name for unrelated tools. At the time of writing, OpenAI's Sora app, website and API have all been shut down. Anything claiming to offer "Sora" now deserves a careful look at who is really behind it.
Believing a demo shows everyday results. Company showcase videos are chosen from many attempts. The Air Head team estimated they kept roughly one second in every 300 they generated. Your first try with any video tool will usually look worse than the demo, and that is normal.
Treating realistic video as proof that the AI understands the world. Sora could make water ripple beautifully and still let a cookie stay whole after someone bit it, an example OpenAI gave itself. Looking right and being physically right are different things; the World Models page goes into this.
Assuming a watermark or label makes AI video safe. Sora's visible watermark could be removed in seconds by free websites, and hidden labels are lost when files are re-encoded. Treat a missing label as "unknown", never as "real".
Thinking an opt-out protects people. Sora's early approach to copyrighted characters put the burden on owners to object, and its handling of historical figures changed only after families complained. With any AI video tool, the safe rule for you is simpler: if you don't have permission to use a person or a character, don't.
Building your work around one company's tool. Developers who built apps on Sora's API had six months' notice to move. Keep copies of what you make, keep your prompts, and prefer workflows you could move to another tool.
Mixing it up with OpenAI's image tools. OpenAI still makes images inside ChatGPT; that is a separate product, covered on the OpenAI Images page. Only the video product, Sora, has closed.
Related elements
- Runway: a video-generation company whose tools, at the time of writing, are still available to creators.
- Veo: Google's video model, a long-running rival to Sora.
- Kling: another AI video generator that competed with Sora.
- Diffusion Model: the technique Sora used to turn noise into moving pictures.
- Transformer: the architecture Sora combined with diffusion to handle video as patches in space and time.
- World Models: the "world simulator" idea OpenAI attached to Sora, and where its research team went next.
- OpenAI Images: OpenAI's image tools in ChatGPT, which use the same kind of provenance labels.
- GPT: OpenAI's main model family, where the company has turned its attention.
- Guardrails: the safety rules Sora kept having to tighten after launch.
- Privacy: why your face and voice are personal data worth protecting.